How to Reduce STL Triangle Count

Simplify an STL only when fewer triangles solve a real handling problem and the resulting surface still meets your requirements. Generate a separate candidate from the original, reopen it and review critical features and mesh findings before choosing it.

Start with a named constraint and feature

There is no starting percentage that protects every model. Use the least geometry change that addresses the receiving constraint, compare the exact saved candidate and keep the original when the required feature cannot be verified. A preset is an input to an experiment, not an acceptance rule.

If the issue is text storage rather than triangles, choose binary encoding without intentional triangle removal. Method selection also includes source re-export and retaining the original.

Owned raised-ridge coupon: a reproducible comparison

Download the original 24-triangle coupon, the 75% target output and the 50% target output. The original has a 40 × 20 × 10 mm block plus a separate touching 16 × 0.6 × 0.5 mm ridge. This deliberately simple two-shell control is not a validated printable solid.

Reproduced in the repository software on 11 October 2026: meshoptimizer 1.2.0, target percentage, relative error 0.01 and LockBorder, then binary export, re-import and collection topology checks. The target is a request, not a promised final count. These observations do not establish that the 75% copy preserves every local detail or that any physical print succeeds.

The meshoptimizer project documents mesh simplification and error/constraint trade-offs; version 1.2 documentation reviewed 11 October 2026. This fixture illustrates why reduction percentage alone is not quality evidence. Simplify a copy locally only when the resulting model meets your own named feature checks.

Running the actual finite comparison on the original and 75% coupon in their original frame, mm on both sides and 0.1 mm tolerance gives 5,000 samples per direction: maximum sampled distance about 0.23290 mm, mean about 0.000512 mm, and 25 of 10,000 samples above tolerance. The largest observed sample is not a whole-surface maximum. A small mean can conceal a meaningful change on a narrow ridge.

The 75% candidate is an observed experiment, not an approval. Inspect the ridge close-up and compare the exact downloaded output in the surface comparison tool. Original coordinates, both millimetre assumptions and tolerance belong in the record. A low global mean does not establish local feature preservation.

Why a requested triangle target can remain unmet

The installed meshoptimizer 1.2.0 simplifier takes a target index count, not a removed-vertex count. InspectMesh computes requested triangles as floor(original triangles × kept percentage / 100), with a minimum of one, and passes three times that count as indices. The owned ridge starts with 24 triangles: 75% requests 18 and produces 18; 50% requests 12 and produces 16. Relative error 0.01 and LockBorder are unchanged.

The version-specific JavaScript contract describes error and topology limits. A relative error flag is an algorithm setting, not a measured physical tolerance or a guaranteed surface bound. Do not disable protections merely to obtain the percentage.

Named small features need their own review

The original narrow-rib control has a 20-coordinate cube, a 2 × 2 × 0.2 marker at (1, 1, 20) and a 2 × 0.1 × 0.2 rib at (9, 9, 20). The diagnostic revision deliberately removes that rib: 36 versus 24 triangles, with the same outer bounds. The marker keeps the maximum height. These touching shells are not a Boolean union or a physical print control; the revision is hand-authored, not a reduction result.

Generate gentler alternatives independently from the original, compare the selected candidate, and retain source CAD when a required feature remains unresolved. A low global mean does not approve a local feature.

Smaller bytes with increased mesh findings

The reproduced 50% ridge output has 16 triangles and 884 bytes, versus 24 and 1,284 bytes in the original, but introduces two duplicate triangles and one over-shared edge. Reject it for the intended print workflow. The reducer retains it for diagnostic review and requires an explicit acknowledgement before a diagnostic download. The original already has two touching shells: that existing ambiguity is distinct from the new findings.

1. Keep the original and reject a candidate with unexplained new findings.
2. Try a gentler independent target and recheck the same supported topology method.
3. Return to source CAD if connectivity or required detail cannot be preserved.
4. Use the repair owner for a separate repair workflow; no silent cleanup occurs during simplification.

Identical finding counts do not prove identical defect locations. Self intersections and physical print success remain unchecked.

Record the candidate decision